A competitive elliptical clustering algorithm

This paper introduces a new learning algorithm for on-line ellipsoidal clus tering. The algorithm is based on the competitive clustering scheme extende d by two specific features. Elliptical clustering is accomplished by effici ently incorporating the Mahalanobis distance measure into the learning rule s, and underutilization of smaller dusters is avoided by incorporating a fr equency-sensitive term. Experiments are conducted to demonstrate the useful ness of the algorithm on artificial data-sets as well as on the problem of texture segmentation. (C) 1999 Elsevier Science B.V. All rights reserved.